We introduce TAPA ( that decouples magnitude from angular contributions in position encoding. TAPA yields better OOD (long-context) performance than vanilla RoPE approach. We also provide theoretical analysis why it works.
Thanks
@yusidwang and the colleagues for the great work!